Accession Number : ADA265637


Title :   An Analog Clustering Network from a Biological Model


Corporate Author : NAVAL COMMAND CONTROL AND OCEAN SURVEILLANCE CENTER RDT AND E DIV SAN DIEGO CA


Personal Author(s) : Shoemaker, P A ; Hutchens, C G ; Patil, S B


Full Text : https://apps.dtic.mil/dtic/tr/fulltext/u2/a265637.pdf


Report Date : Apr 1993


Pagination or Media Count : 4


Abstract : One reason for the recent resurgence of interest in neural network- like computational models has been the prospect of compact and fast implementations of these networks in integrated circuit form. While analog implementation offe considerable advantages with regard to speed and density, their precision and noise immunity are important concerns. Some researchers (e. g., Mead and coworkers) have built analogues of biological structures for early sensory processing, and they have emphasized that tolerance of noisy and imprecise components is a natural emergent feature of these networks. However, the ways in which higher or cognitive functions might be learned and computed with such components for the most part remains unknown. In addition, learning itself remains problematic in analog circuitry. Means proposed for long-term, modifiable analog weight storage (e.g., floating-gate MOS devices) are sensitive, difficult to control, and of limited precision. We have chosen to implement a model of olfactory processing proposed by Granger, Lynch and Ambros- Ingerson, which we believe to be an instructive paradigm for computation in a learning system with low-precision weights and weight changes. The model has been shown capable of performing a hierarchical clustering of vectors on its input space. This capability is of potential interest for a range of applications, from automatic target recognition (ATR) to surveillance and detection. The network requires only coarse-valued weights (three to five bits resolution) and its operation relies on the statistical properties of large assemblages of sparsely interconnected neurons, rather than high precision processing.


Descriptors :   *NEURAL NETS , *INTEGRATED CIRCUITS , *ARTIFICIAL INTELLIGENCE , SIGNAL PROCESSING , COMPUTERIZED SIMULATION , COMPUTATIONS , DETECTION , COGNITION , RESOLUTION , TARGETS , SURVEILLANCE , STORAGE , NOISE REDUCTION , NOISE , NERVE CELLS , RECOGNITION , COMPUTER APPLICATIONS , PRECISION , CLUSTERING , SMELL , OLFACTORY NERVE , LEARNING MACHINES , STRUCTURES , STATISTICAL DATA , MODELS , ENVIRONMENTS , INPUT , DENSITY


Subject Categories : Cybernetics


Distribution Statement : APPROVED FOR PUBLIC RELEASE